Editorial illustration for Liquid AI Builds Hybrid Models With 80% Liquid Neural Networks
Liquid AI's 80% Liquid Neural Networks Challenge LLM Design
Liquid AI Builds Hybrid Models With 80% Liquid Neural Networks
Liquid AI, a Boston-based startup spun out of MIT, is building language models where roughly 80% of the architecture runs on liquid neural networks instead of standard transformer blocks. That ratio matters because it marks a break from the design that has powered nearly every major LLM since Google researchers published "Attention Is All You Need" in 2017. Transformers gave the industry GPT, Claude, Gemini and everything after.
Nine years later, cracks are showing. Reasoning models and long-context tricks that dominate current headlines are largely patches on a transformer core that struggles with efficiency at scale, not clean upgrades to it.
Liquid AI is one of several startups betting that the next wave of progress requires rethinking that core rather than bolting on more workarounds. Justin Dangel, cofounder and CEO of rival startup Subquadratic, calls transformers "one of the most important innovations in the history of computer science." The question dividing this new crop of companies isn't whether transformers changed the world. It's whether hybrid architectures, blending transformer components with alternatives like liquid neural networks, can outrun the limits that are starting to show.
A growing number of scientists and engineers are now asking what’s coming next. LLMs are not going anywhere, but the way they get built is up for grabs.
Why this matters
Liquid AI's bet is that transformers, the architecture behind nearly every major LLM since 2017, aren't the endpoint. An 80/20 split toward liquid neural networks, tuned by an in-house designer AI rather than human intuition, suggests the company is treating architecture search itself as the product. That's worth watching closely if you're building on top of foundation models: a startup that automates its own model design could iterate on efficiency gains faster than teams hand-tuning transformer variants.
Ramin Hasani calling this "the core technology of our company" is a signal about priorities, not proof of results yet. We haven't seen benchmarks here, so skepticism is warranted until Liquid AI publishes numbers on inference cost, latency, or accuracy against comparable transformer-based models. For founders weighing infrastructure bets, the real question is whether hybrid architectures like this deliver enough efficiency to justify moving off the transformer ecosystem everyone else has spent eight years optimizing around.
That answer isn't in yet, but the approach itself, letting an AI design the AI, is the part worth tracking.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives ByteDance's SeedRealtime LLM Watches, Listens and Speaks
From Archives Google's DiffusionGemma Solves 85% of Puzzles After Minimal Fine-Tuning
Common Questions Answered
What percentage of Liquid AI's hybrid models use liquid neural networks instead of transformer blocks?
Liquid AI's hybrid models are composed of approximately 80% liquid neural networks, with the remaining 20% using standard transformer blocks. This 80/20 split represents a significant departure from the transformer-based architecture that has dominated LLM design since Google's 2017 'Attention Is All You Need' paper.
How does Liquid AI's approach to architecture design differ from traditional LLM development?
Rather than relying on human intuition to design model architectures, Liquid AI uses an in-house designer AI to tune its hybrid models. This approach treats architecture search itself as a product, potentially allowing the company to iterate on efficiency gains faster than teams using manual design methods.
Why is Liquid AI challenging the transformer architecture that has powered major LLMs like GPT and Claude?
According to the article, cracks are showing in the transformer design that has been standard since 2017, particularly in reasoning models and long-context tasks. Liquid AI's hybrid approach suggests that transformers may not be the endpoint of LLM architecture evolution, and the company believes liquid neural networks can address current limitations.
What is the significance of Liquid AI being spun out of MIT?
Liquid AI's MIT origins provide the startup with access to cutting-edge research and talent in neural network design and AI architecture. This academic foundation supports the company's ambitious goal of developing alternative architectures to the transformer-based models that have dominated the industry for nearly a decade.
Further Reading
- Liquid Foundation Models - Liquid AI
- From Liquid Neural Networks to Liquid Foundation Models - Liquid AI
- MIT offshoot Liquid AI releases blueprint for enterprise-grade small-model training - VentureBeat
- Liquid AI Released LFM2.5-350M: A Compact 350M Parameter Model Trained on 28T Tokens With Scaled Reinforcement Learning - MarkTechPost
- Hybrid transformer model with liquid neural networks and learnable ... - ScienceDirect
